[chrisfosterelli]

I was recently talking to a colleague I went to school with and they said the same thing, but for a different reason. We both did grad studies with a focus on ML, and at the time ML as a field seemed to be moving so fast. There was a lot of excitement around AI again finally after the 'AI winter'. It was easy to participate in bringing something new to the field, and there was so many unique and interesting models coming about every day. There was genuine discussion about a viable path to AGI.

Now, basically every new "AI" feature feels like a hack on top of yet another LLM. And sure the LLMs seem to keep getting marginally better, but the only people with the resources to actually work on new ones anymore are large corporate labs that hide their results behind corporate facades and give us mere mortals an API at best. The days of coding a unique ML algorithm for a domain specific problem are pretty much gone -- the only thing people pay attention to is shoving your domain specific problem into an LLM-shaped box. Even the original "AI godfathers" seem mostly disinterested in LLMs these days, and most people in ML seem dubious that simply scaling up LLMs more and more will be a likely path to AGI.

It seems like there's more excitement around AI for the average person, which is probably a good thing I suppose, but for a lot of people that were into the field they're not really that fun anymore.

In terms of programming, I think they can be pretty fun for side projects. The sort of thing you wouldn't have had time to do otherwise. For the sort of thing you know you need to do anyway and need to do well, I notice that senior engineers spend more time babysitting them than benefitting from them. LLMs are good at the mechanics of code and struggle with the architecture / design / big picture. Seniors don't really think much about the mechanics of code, it's almost second nature, so they don't seem to benefit as much there. Juniors seem to get a lot more benefit because the mechanics of the code can be a struggle for them.

[porker]

> Now, basically every new "AI" feature feels like a hack on top of yet another LLM.

LLM user here with no experience of ML besides fine-tuning existing models for image classification.

What are the exciting AI fields outside of LLMs? Are there pending breakthroughs that could change the field? Does it look like LLMs are a local maxima and other approaches will win through - even just for other areas?

Personally I'm looking forward to someone solving 3D model generation as I suck at CAD but would 3D print stuff if I didn't have to draw it. And better image segmentation/classification models. There's gotta be other stuff that LLMs aren't the answer to?

[chrisfosterelli]

Well one of the inherent issues is assuming that text is the optimal modality for every thing we try to use an LLM for. LLMs are statistical engines designed to predict the most likely next token in a sequence of words. Any 'understanding' they do is ultimately incidental to that goal and once you look at them that way a lot of the shortcomings we see become more intuitive.

There's a lot of problems LLMs are really useful for because generating text is what you want to do. But there's tons of problems which we would want some sort of intelligent, learning behaviour that do not map to language at all. There's also a lot of problems that can "sort of" be mapped to a language problem but make pretty extraneous use of resources compared to a (existing or potential) domain specific solution. For purposes of AGI, you could argue that trying to express "general intelligence" via language alone is fundamentally flawed altogether -- although that quickly becomes a debate about what actually counts as intelligence.

I pay less attention to this space lately so I'm probably not the most informed. Everyone seems so hyped about LLMs that I feel like a lot of other progress gets buried, but I'm sure it's happening. There's some problem domains that are obviously solved better with other paradigms currently: self-driving tech, recommendation systems, robotics, game AIs, etc. Some of the exciting stuff that can likely solve some problems better in the future is some of the work on world models, graph neural nets, multi modality, reinforcement learning, alternatives to gradient descent, etc. I think it's a debate whether or not LLMs are a local maxima but many of the leading AI researchers seem to think so -- Yann Lecun recently for e.g. said LLMs 'are not a path to human-level AI'

[jiggawatts]

It’s now moving faster than ever. Huge strides have been made in interpretability, multi modality, and especially the theoretical understanding of how training interacts with high dimensional spaces. E.g.: https://transformer-circuits.pub/2022/toy_model/index.html

[chrisfosterelli]

Thanks, this seems interesting. I'll give it a read. I admittedly don't keep tabs as much as I should these days. I feel like every piece of AI news is about LLMs. I suppose I should know other people are still doing interesting things :)